CAREER: Fast, Accurate Estimation and Prediction using Markov Logic

University Of Texas At Dallas, Richardson TX

Investigators

Abstract

Markov logic networks (MLNs) are routinely used in a wide variety of application domains including information extraction, computer vision, bio-informatics, and natural language understanding to represent and reason about relational and probabilistic knowledge. However, inference and learning in them is extremely challenging and despite tremendous progress in recent years several key real-world reasoning tasks remain out of reach. The goal of this CAREER award is to vastly improve the scalability and accuracy of learning and inference algorithms for MLNs and thus solve much larger and harder reasoning problems than is possible today. The award also includes a tightly integrated education and outreach component. Specifically, it (1) involves high school students as well as undergraduate students, especially young women in development and model building exercises, encouraging them to pursue career in research; (2) yields open source software tools to facilitate and broaden the adoption of MLN technology; and (3) promotes standardization of datasets and evaluation methodologies via organization of inference competitions. The key technical contribution of the proposed research is to address two fundamental limitations of lifted probabilistic inference algorithms, the dominating approach for inference and learning in MLNs. First, existing lifted methods primarily exploit exact symmetries and ignore approximate symmetries. Second, MLNs having large number of symmetries are often not expressive enough to accurately model complex, real-world dependencies and hidden phenomena. The CAREER award addresses these issues by developing: (1) principled approaches that exploit approximate symmetries; (2) novel learning algorithms that use structured latent variables to induce diverse, highly expressive MLNs; and (3) a unifying message-passing framework called lifted structured message passing that systematically exploits approximate symmetries and structured representations for solving a range of inference tasks defined over MLNs including marginal estimation, maximum-a-posteriori estimation and marginal maximum-a-posteriori estimation.